A bug occurs when there are few explanatory variables

[mamitsu2]

There are few explanatory variables, so bug occurs. Please fix it!

[myamada0321]

Thanks for the comments.

Could you share the dataset you used and your setup? Then, we can reproduce the bug.

Thanks!

[mamitsu2]

Sorry to be inadequate. I uses sklearn.datesets.load_boston to test this module. If I do as follows, bug doesn't occur.

dataset = load_boston()
# set dataframe
X1_ = pd.DataFrame(dataset.data, columns=dataset.feature_names)
y1_ = pd.DataFrame(dataset.target, columns=['y'])
X1_ = X1_.iloc[:,:]
X1 = np.array(X1_)
y1 = np.array(y1_)
X1_col = X1_.columns
hsic_lasso = HSICLasso()
hsic_lasso.input(X1,y1.flatten(),featname=list(X1_col))
hsic_lasso.regression(num_feat=X1.shape[1], discrete_x=False, n_jobs=2)
hsic_lasso.dump()
hsic_lasso.get_index_score()

but, I do as follows, reduce explanatory variables,

X1_ = X1_.iloc[:,:5]

ValueError: attempt to get argmax of an empty sequence is occured.

A bug occurs here.

~/.pyenv/versions/anaconda3-5.2.0/lib/python3.6/site-packages/pyHSICLasso/nlars.py in nlars(X, X_ty, num_feat, max_neighbors)
     97         XtXbeta = np.dot(X.transpose(), np.dot(X, beta))
     98         c = X_ty - XtXbeta
---> 99         j = np.argmax(c[I])
    100         C = max(c[I])
    101 

[myamada0321]

Thanks for the detailed information. We will investigate this case.

[muoten]

I've got the same problem with another dataset with few explanatory variables.

Then I've replicated the error with sklearn.datasets.load_boston and 5 features. It seems that I array gets empty ([]) when lasso_cond=0. And this exception is not controlled on the while loop or compensated anyway.

Any hint to fix this issue? I think the library is very interesting, and HSIC-based optimization may be useful too for datasets with few columns.

Thank you!

[hclimente]

Thanks for your input. We have been looking at alternative Lasso solvers. Unfortunately, we haven't found one that checks all of our boxes... We'll be on the lookout for a new solvers that would address this issue.

[renero]

Thanks for reporting the bug. Has anyone solved it?